mischeiwiller/swiss-german-text

Swiss German Text (Mundart)

An open corpus of written Swiss German (gsw). Written Swiss German is a spoken-first, low-resource variety with little clean, openly-licensed, machine- readable text. This is the v1 release of the Mundart project: 103,462 clean, schema-validated, PII-scrubbed rows of written Swiss German / Alemannic drawn from three greenlit, openly-licensed sources — the Alemannic Wikipedia, the Tatoeba gsw sentence export, and the Leipzig Corpora Collection gsw web slice — packaged as a single Parquet file under one uniform row schema.

ℹ️ Status: v1 — three greenlit sources. This is a defensible open written Swiss German corpus combining an encyclopedic anchor, a small parallel (gsw→de) slice, and a large web-sentence slice. It is not a balanced or representative census of the language — see Limitations. The mix is broader Alemannic, not pure Swiss German.

Dataset at a glance

Rows	103,462
Language	Swiss German / Alemannic (gsw)
Sources	3 (Alemannic Wikipedia + Tatoeba gsw + Leipzig gsw web)
Region labels	unknown (not source-derived — see below)
Parallel High German (de_aligned)	334 rows (Tatoeba gsw→deu links)
Format	Parquet (single file)
Licenses	CC-BY-SA-4.0 / CC-BY-2.0-FR / CC-BY-4.0 (per row, see below)

Sources

Source (source)	Rows	License (license)	Provenance
als-wikipedia	2,940	CC-BY-SA-4.0	Alemannic Wikipedia, Wikimedia dump 20260601
tatoeba-gsw	523	CC-BY-2.0-FR	Tatoeba gsw sentence export, 20260606 (334 with de_aligned)
leipzig-gsw-web	99,999	CC-BY-4.0	Leipzig Corpora Collection, package gsw-ch_web_2017_100K
Total	103,462

Each row carries its own source and license; the published corpus is the union of three CC-BY-family licenses (all ShareAlike-compatible). The ShareAlike obligation of the Alemannic-Wikipedia slice means the combined dataset is offered under CC-BY-SA-4.0.

Intended use

• Low-resource NLP — language modelling, masked-LM pre-training, and domain-adaptation for written Swiss German / Alemannic.
• Dialect identification — training/eval signal for gsw-vs-de and Alemannic dialect-ID tasks (the companion mischeiwiller/gsw-eval).
• Linguistic research — an openly-licensed, reproducible written-gsw reference corpus.

Out of scope / non-goals

• Not a balanced dialect sample: regions are not labelled (all unknown).
• Not primarily parallel data: only 334 rows (all Tatoeba) carry aligned High-German text; the rest are monolingual.
• Not pure Swiss German: the mix is broader Alemannic (see Limitations).
• No verbatim re-publication beyond what the per-row CC-BY-family license permits.

Schema

One row = one cleaned text unit (a Wikipedia article, or a single sentence for the Tatoeba / Leipzig slices).

Field	Type	Notes
text	string	Cleaned, markup-stripped, PII-scrubbed Swiss German / Alemannic text.
source	string	Provenance id of the source slice (als-wikipedia / tatoeba-gsw / leipzig-gsw-web).
license	string	License of the source slice (CC-BY-SA-4.0 / CC-BY-2.0-FR / CC-BY-4.0).
region	string	Coarse dialect region, or unknown. All unknown (not source-derived).
de_aligned	string | null	Parallel High-German rendering, or null. Populated only for the 334 linked Tatoeba rows.
meta	struct	Per-source provenance; the union of all slices' keys ({title, lang, dump_date, dump_url, sentence_id, export_date, package, source_url}). A row leaves the keys its source doesn't supply as null.

The canonical row schema is published as JSON Schema in the project repo (schema/row.schema.json) and enforced by pydantic at ingest + validation.

How it was built

Each slice is ingested by its own runner, normalized to the shared row schema, PII-scrubbed, validated, then concatenated into one Parquet file:

1. Alemannic Wikipedia (als-wikipedia) — stream the pinned alswiki-20260601-pages-articles.xml.bz2 Wikimedia dump; keep main-namespace, non-redirect articles; strip MediaWiki markup with mwparserfromhell (strip_code, normalized + collapsed); drop blank lines; drop articles whose cleaned text is under 200 characters (stubs).
2. Tatoeba gsw (tatoeba-gsw) — parse the pinned 20260606 gsw sentence export; populate de_aligned only from genuine gsw→deu translation links (the smallest resolving deu id, never inferred); drop one- and two-token greetings.
3. Leipzig gsw web (leipzig-gsw-web) — stream the gsw-ch_web_2017_100K package's *-sentences.txt; preserve the upstream shuffled + deduplicated order; monolingual (de_aligned null).
4. PII scrub — see PII handling below (applied to text and to de_aligned).
5. Validate — every row is validated against the CorpusRow pydantic model and the committed JSON Schema; seeded samples are re-audited for PII leaks (each slice: 0 leaks).
6. Package — write a single typed Parquet file under an explicit Arrow schema mirroring the row schema; re-scrub idempotently at packaging time.

Reproduction code: https://github.com/lappemic/hf-lab (project 03-mundart).

PII handling

Contact PII is redacted to typed placeholders by an ordered, idempotent regex pass (mundart.clean.scrub_pii), applied before the length gate at ingest:

• emails → [EMAIL]
• URLs → [URL]
• social handles → [HANDLE]
• phone numbers → [PHONE] (conservative: international +CC runs and Swiss national numbers that carry at least one separator; bare digit runs are treated as catalogue/DB ids, not phones).

Person names are deliberately NOT redacted. The sources are public, already-published text (encyclopedic articles, community sentences, web sentences); blanket name redaction would gut the corpus and is not required. Only contact PII is removed. The scrub runs on every slice's text (and de_aligned); a seeded post-scrub audit on each slice found 0 leaks.

Region-label provenance

region is unknown for every row. None of the three sources carries reliable, per-row dialect-region metadata, and we do not infer a normative dialect label from text. A region is only ever populated when a source supplies it directly. Treating region as ground truth would be incorrect.

Provenance & legal

Each row's exact provenance (dump/export date, package, source URL, sentence id) is stamped in its meta struct. By slice:

• Alemannic Wikipedia (als-wikipedia, 2,940 rows) — als.wikipedia.org Wikimedia dump 20260601 (alswiki-20260601-pages-articles.xml.bz2, https://dumps.wikimedia.org/alswiki/20260601/). License CC-BY-SA-4.0.
• Tatoeba gsw (tatoeba-gsw, 523 rows, 334 with de_aligned) — the gsw per-language sentence export, download date 20260606 (https://downloads.tatoeba.org/exports/per_language/). License CC-BY-2.0-FR (the detailed export carries no per-sentence license column; the dataset-wide CC-BY-2.0-FR is applied conservatively).
• Leipzig Corpora Collection (leipzig-gsw-web, 99,999 rows) — the non-Wikipedia web package gsw-ch_web_2017_100K (https://downloads.wortschatz-leipzig.de/corpora/). The downloadable text corpora are offered under CC-BY (distinct from the NC online query interface); stamped CC-BY-4.0.

Combined license: the Alemannic-Wikipedia slice is ShareAlike, so the combined dataset is offered under CC-BY-SA-4.0. CC0, CC-BY-2.0-FR and CC-BY are all one-way inbound-compatible into CC-BY-SA-4.0 — ShareAlike binds adapted material, not an aggregation of separately-licensed sentences. Only sources explicitly greenlit in the project's source ledger are ingested; NC/ND/research-only/unclear sources are excluded.

Attribution (required, per source — text was extracted, sentence-filtered and concatenated):

• Alemannic Wikipedia — © Wikipedia contributors, Alemannische Wikipedia, CC-BY-SA-4.0, https://als.wikipedia.org.
• Tatoeba — © Tatoeba contributors, CC-BY-2.0-FR (the CC0 subset is conservatively over-attributed as BY), https://tatoeba.org.
• Leipzig Corpora Collection — the gsw-ch web download package, CC-BY. The downloadable corpora are CC-BY — distinct from the CC-BY-NC online query interface (so a third-party HF mirror tagged "NC" is not the download license). Cite Goldhahn, Eckart & Quasthoff, Building Large Monolingual Dictionaries at the Leipzig Corpora Collection, LREC 2012. https://wortschatz.uni-leipzig.de.

Limitations

• Broader Alemannic, not pure Swiss German. The Alemannic-Wikipedia slice mixes Swiss German with Alsatian, Swabian and other Alemannic varieties; the Leipzig web slice is gsw-ch but still uncurated for dialect. We cannot defensibly split these per row, so region stays unknown and downstream gsw-only consumers should account for the dialectal spread.
• Source-skewed. ~97% of rows come from the Leipzig web slice (single sentences), ~3% from Wikipedia (articles), and a small parallel tail from Tatoeba. Genre, length and register are not balanced across sources.
• Sentence- vs document-level mix. Wikipedia rows are full articles; Tatoeba and Leipzig rows are single sentences. text length varies by orders of magnitude across sources.
• Sparse parallel data. Only 334 rows carry de_aligned (all Tatoeba); the corpus is overwhelmingly monolingual.
• No region labels (region all unknown), by design — never inferred.
• Residual markup. strip_code is robust but not perfect; rare template/ table artefacts may survive in the Wikipedia slice.

Companion artifacts

Part of the Mundart project (dataset + eval + demo):

• Eval: mischeiwiller/gsw-eval — gsw-vs-de LID task + baseline, derived from this corpus.
• Space: mischeiwiller/mundart-explorer — paste Swiss German → gsw-vs-de LID guess + High-German rendering where available.
• Neighbour reference: ZurichNLP/swissbert — a Swiss-German-aware multilingual encoder; a natural model to pre-train or fine-tune on this corpus.

Citation

@misc{mundart_swiss_german_2026,
  title        = {Swiss German Text (Mundart)},
  author       = {Scheiwiller, Michael},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  note         = {Derived from the Alemannic Wikipedia (dump 20260601),
                  the Tatoeba gsw export (20260606) and the Leipzig Corpora
                  Collection (gsw-ch_web_2017_100K). CC-BY-SA-4.0.}
}